Compiler Optimizations for a Time-constrained Environment

93 Pages·2008·0.54 MB·English

by  Owen Anderson

Compiler Optimizations for a Time-constrained Environment
Owen C. Anderson
Prof. Susan Fox, First reader
Prof. Libby Shoop, Second reader
Prof. Andrew Beveridge, Third reader
April,2008
DepartmentofMathematicsandComputerScience
Copyright c 2008OwenC.Anderson.
!
TheauthorgrantsMacalesterCollegethenonexclusiverighttomakethisworkavailablefornoncommer-
cial,educationalpurposes,providedthatthiscopyrightstatementappearsonthereproducedmaterialsand
notice is given that the copying is by permission of the author. To disseminate otherwise or to republish
requireswrittenpermissionfromtheauthor.
Abstract
Overthelastseveraldecades,twoimportantshiftshavetakenplaceinthecomputingworld: first,
thespeedofprocessorshasvastlyoutstrippedthespeedofmemory,makingmemoryaccessesby
farthemostexpensiveoperationsthatatypicalsymbolicprogramperforms. Second,dynamically
compiledlanguagessuchasJavaandC#havebecomepopular,placingnewpressuresoncompiler
writerstocreateeffectivesystemsforrun-timecodegeneration.
Thispaperaddressestheneedcreatedbythelaggingspeedsofmemoryaccessesinthecontext
of dynamically compiled systems. In such systems memory access optimization is important for
resultant program performance, but the compilation time required by most traditional memory
access optimizations is prohibitively high for use in such contexts. In this paper, we present a
newanalysis,memorydependenceanalysis,whichamortizesthecostofperformingmemoryaccess
analysistoalevelthatisacceptablefordynamiccompilation. Inaddition,wepresenttwomemory
access optimizations based on this new analysis, and present empirical evidence that using this
approachresultsinsignificantlyimprovedcompilationtimeswithoutsignificantlossinresultant
codequality.

Contents
Abstract iii
Contents v
ListofFigures vii
ListofTables ix
Acknowledgments xi
1 Introduction 1
1.1 ModernStaticOptimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 EvolutionofJust-in-TimeCompilation . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 PreviousWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Implementation: LLVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 AliasAnalysis 7
2.1 ClassicalAliasAnalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 BasicAliasAnalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 MemoryDependenceAnalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3 DeadStoreElimination 13
3.1 ClassicalDeadStoreElimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 FastDeadStoreElimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4 RedundantLoadElimination 19
4.1 GlobalCommonSubexpressionElimination. . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 GlobalValueNumbering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
vi Contents
5 Conclusions 25
5.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 FutureWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Bibliography 29
CodeListing 31
List of Figures
1.1 Anexamplecontrol-flowgraph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 AsimpleexampleinCandLLVMIR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Anexampleofmemorydependenceanalysis . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Adiagramaticexampleoftheinternalsofmemorydependenceanalysis . . . . . . . 10
3.1 Anexamplewherethefirststoreispotentiallynecessary . . . . . . . . . . . . . . . . 13
3.2 Anexampleinwhichthefirststoreisdead . . . . . . . . . . . . . . . . . . . . . . . . 14
4.1 Anexamplewheretheloadisnotredundant . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Anexamplewheretheloadisredundant . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Theresultofeliminatingaredundantload . . . . . . . . . . . . . . . . . . . . . . . . 21
5.1 Totaltimetoexecutetheoptimizationsinsecondsofthefourlargesttestcasesfrom
SPEC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.2 NormalizedexecutiontimeofthefourlargesttestcasesfromSPEC . . . . . . . . . . 26

List of Tables
2.1 ThenumberofaliasqueriesevaluatedontheSPECbenchmarks . . . . . . . . . . . . 11
3.1 ThenumberofstoresremovedontheSPECbenchmarks . . . . . . . . . . . . . . . . 16
4.1 ThenumberofloadsremovedontheSPECbenchmarks . . . . . . . . . . . . . . . . 23